|Publication number||US7908275 B2|
|Application number||US 10/590,397|
|Publication date||Mar 15, 2011|
|Filing date||Jul 3, 2006|
|Priority date||Jul 3, 2006|
|Also published as||CN101553799A, CN101553799B, EP2044524A1, EP2044524A4, US20090019025, US20110184952, WO2008006241A1|
|Publication number||10590397, 590397, PCT/2006/1550, PCT/CN/2006/001550, PCT/CN/2006/01550, PCT/CN/6/001550, PCT/CN/6/01550, PCT/CN2006/001550, PCT/CN2006/01550, PCT/CN2006001550, PCT/CN200601550, PCT/CN6/001550, PCT/CN6/01550, PCT/CN6001550, PCT/CN601550, US 7908275 B2, US 7908275B2, US-B2-7908275, US7908275 B2, US7908275B2|
|Original Assignee||Intel Corporation|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (19), Non-Patent Citations (9), Referenced by (1), Classifications (9), Legal Events (2)|
|External Links: USPTO, USPTO Assignment, Espacenet|
This U.S. application claims priority to Pending International Application Number PCT/CN2006/001550, filed PCT in China on Jul. 3, 2006.
This disclosure relates generally to signal processing and multimedia applications, and more specifically but not exclusively, to methods and apparatus for fast audio search and audio fingerprinting.
Audio search (e.g., searching a large audio stream for an audio clip, even if the large audio stream is corrupted/distorted) has many applications including analysis of broadcast music/commercials, copyright management over the Internet, or finding metadata for unlabeled audio clips, and etc. A typical audio search system is serial and designed for single processor systems. It normally takes a long time for such a search system to search for a target audio clip in a large audio stream. In many cases, however, an audio search system is required to work efficiently on large audio databases, e.g., to search large databases in a very short time (e.g., close to real-time). Additionally, an audio database may be partially or entirely distorted, corrupted, and/or compressed. This requires that an audio search system be robust enough to identify those audio segments that are the same as the target audio clip, even if those segments may be distorted, corrupted, and/or compressed. Thus, it is desirable to have an audio search system which can quickly and robustly search large audio databases for a target audio clip.
The features and advantages of the disclosed subject matter will become apparent from the following detailed description of the subject matter in which:
According to embodiments of the subject matter disclosed in this application, a large audio stream or a large audio database in a multiprocessor system may be searched for a target audio clip using a robust and parallel search method. The large audio database may be partitioned into a number of smaller groups. These smaller groups may be dynamically scheduled to be processed by available processors or processing cores in the multiprocessor system. Processors or processing cores may process the scheduled groups in parallel by partitioning each group into smaller segments, extracting acoustic features from the segments; and modeling the segments using a common component Gaussian Mixture model (“CCGMM”). The length of these segments may be the same as the length of the target audio clip. Before processing any group, one processor or processing core may extract acoustic features from the target audio clip and model it using the CCGMM. A Kullback-Leibler (KL) or KL-max distance may be further computed between the model of the target audio clip and each segment of a group. If the distance equals or smaller than a predetermined value, the corresponding segment is identified as the target audio clip.
If the distance is larger than a predetermined value, the processor or processing core may skip a certain number of segments and continue searching for the target audio clip. Once a processor or processing core finishes searching a group, a new groups may be given to it for processing to search for the target audio clip until all of the groups are searched. The size of the groups may be determined in such a way to reduce the load imbalance and the overlapped computation. Furthermore, Input/Output (I/O) may be optimized to improve the efficiency of parallel processing of audio groups by multiple processors or processing cores.
Reference in the specification to “one embodiment” or “an embodiment” of the disclosed subject matter means that a particular feature, structure or characteristic described in connection with the embodiment is included in at least one embodiment of the disclosed subject matter. Thus, the appearances of the phrase “in one embodiment” appearing in various places throughout the specification are not necessarily all referring to the same embodiment.
The audio search module may first partition a large audio database into multiple smaller groups or a large audio stream into smaller partially overlapped substreams. Second, one core may process an audio clip to be searched for (“target audio clip”) to establish a model for the target audio clip. In the mean while, the audio search module dynamically schedules smaller audio groups/substreams to multiple cores, which partition each group/substream into segments and establish a model for each audio segment, in parallel. The size of each segment may be equal to the size of the target audio clip. A Gaussian mixture model (“GMM”) with multiple Gaussian components, which are common to all of the audio segments including both the target audio clip and the audio database/stream, may be used for modeling each audio segment and the target audio clip. Once a model is established for an audio segment, Kullback-Leibler (“KL”) or KL-max distance may be computed between the segment model and the target audio clip model. If the distance is not larger than a predetermined value, the audio segment may be identified as the target audio clip. The search process may continue until all audio groups/substreams are processed.
The computing system 100 may also include a chipset 130 coupled to the system interconnect 115. Chipset 130 may include one or more integrated circuit packages or chips. Chipset 130 may comprise one or more device interfaces 135 to support data transfers to and/or from other components 160 of the computing system 100 such as, for example, BIOS firmware, keyboards, mice, storage devices, network interfaces, etc. Chipset 130 may be coupled to a Peripheral Component Interconnect (PCI) bus 170. Chipset 130 may include a PCI bridge 145 that provides an interface to the PCI bus 170. The PCI Bridge 145 may provide a data path between the processor 110 as well as other components 160, and peripheral devices such as, for example, an audio device 180 and a disk drive 190. Although not shown, other devices may also be coupled to the PCI bus 170.
Additionally, chipset 130 may comprise a memory controller 125 that is coupled to a main memory 150. The main memory 150 may store data and sequences of instructions that are executed by multiple cores of the processor 110 or any other device included in the system. The memory controller 125 may access the main memory 150 in response to memory transactions associated with multiple cores of the processor 110, and other devices in the computing system 100. In one embodiment, memory controller 150 may be located in processor 110 or some other circuitries. The main memory 150 may comprise various memory devices that provide addressable storage locations which the memory controller 125 may read data from and/or write data to. The main memory 150 may comprise one or more different types of memory devices such as Dynamic Random Access Memory (DRAM) devices, Synchronous DRAM (SDRAM) devices, Double Data Rate (DDR) SDRAM devices, or other memory devices.
The overlap between two adjacent segments is to reduce the likelihood of missing any target audio clip between two adjacent segments. The longer the overlap is, the less likely a miss is. In one embodiment, the overlap may be equal to the length of a segment minus the length of a frame to avoid missing any match. However, longer overlap means more computation. Thus, there should be a balance between the computation load and the likelihood of miss (e.g., the overlap is equal to or less than ½ of the segment length). In any case, feature vectors for frames that are overlapped between two segments only need to be extracted once.
Modeling mechanism 420 may establish a model for an audio segment based on its feature vector sequence extracted by feature extractor 410. Depending on what model is used, the modeling mechanism will estimate parameters for the model. In one embodiment, a common component Gaussian mixture model (“CCGMM”) may be used for modeling an audio segment. The CCGMM includes multiple Gaussian components which are common across all of the segments. For each segment, the modeling mechanism estimates a specific set of mixture weights for the common Gaussian components. In another embodiment, other models (e.g., hidden Markov model) may be used for modeling an audio segment. In one embodiment, only the target audio clip may be modeled; and the feature vector sequence of an audio segment may be directly used to determine whether the audio segment is substantially the same as the target audio clip.
Decision maker 430 may determine whether an audio segment in the input audio stream is sufficiently similar so that the audio segment can be identified as a copy of the target audio clip. To achieve this goal, the decision maker may derive a similarity measure by comparing the model of the audio segment and the model of the target audio clip. In one embodiment, the similarity measure may be a distance computed between the two models. In another embodiment, the similarity measure may be probability of the audio segment model being the same as the target audio clip model. Yet in another embodiment, the similarity measure may be derived by comparing the feature vector sequence of the audio segment and the model of the target audio clip. For example, when a hidden Markov model (“HMM”) is used to model the target audio clip, a Viterbi based algorithm may be used to compute a likelihood score between the audio segment and the target audio clip, based on the feature vector sequence of the audio segment and the HMM of the target audio clip.
Based on the value of the similarity measure, the decision maker may determine whether an audio segment can be identified as the target audio clip. For example, if the value of the similarity measure is not larger than a predetermined threshold (e.g., similarity measure is distance between the audio segment model and the target audio clip), the audio segment may be identified as substantially the same as the target audio clip. Similarly, the audio segment may be identified as substantially the same as the target audio clip if the value of the similarity measure is not smaller than a predetermined threshold (e.g., similarity measure is a likelihood score of the audio segment being substantially the same as the target audio clip). On the other hand, if an audio segment is found to be substantially different from the target audio clip based on the similarity measure, a certain number of segments immediately following the audio segment may be skipped. The actual number of segments to be skipped will depend on the value of the similarity measure and/or empirical data. By skipping a number of following segments, it is not likely to miss any target audio clip when the similarity measure indicate the current segment is so different from the target audio clip because the window used to partition an input audio stream into segments slides forward gradually and as a result, there is continuity of similarity measure from one segment to the next.
The GMM, P(k)(x), includes M Gaussian components with component weights wi (k), means μi (k), and covariance
with i=1, 2, . . . , M; wherein k denotes segment k and N( ) denotes a Gaussian distribution. For the target audio clip, there is only one segment, and hence there is no need to use k to identify a segment. For the input audio stream 520, however, there is typically more than one segment, and it is thus desirable to identify the GMM for different segments.
In the example shown in
As a result, Equation (1) becomes:
For each audio segment, only a set of weights, Wi (k), i=1, 2, . . . , M, need to be estimated for the common Gaussian components. Given a feature vector sequence for segment k, which has T feature vectors, xi(t=1, 2, . . . , T), weights may be estimated as follows,
wherein wi (μ) or wj (μ) is a universal weight for the ith or jth segment, which may be obtained by experiments based on some sample audio files or be initialized with a random value.
An input audio stream 520, which is to be searched for the target audio clip 510, may be received by a feature extractor. At block 530B, the feature extractor partitions the input audio stream into partially overlapped segments. For each segment, the feature extractor further partitions the segment into multiple partially overlapped frames and extracts a feature vector from each frame. Block 560 shows a feature vector sequence for the input audio stream 520 and also illustrates how the audio stream is partitioned into partially overlapped segments. For example, a window with the size being the same as the length of the target audio clip may be applied to input audio stream 520. For illustration purpose, a window is shown for the feature vector sequence of the target audio clip to obtain a segment 560A although there is typically no need to apply a window to the target audio clip since there is only one segment. A shifting window is applied to the input audio stream to obtain multiple partially overlapped segments such as 560B and 560C. The window shifts by time τ from segment 560B to segment 560C, where r is smaller than the window size.
Each audio segment is modeled using the CCGMM, for example, segment 560B is modeled at block 570B and segment 560C is modeled at block 570C. Models for each segment of input audio stream 520 and for target audio clip 510 have common Gaussian components with different sets of weights. In one embodiment, feature vectors may be extracted from the entire input audio stream frame by frame to produce a long feature vector sequence for the entire input audio stream. A window with a length being N×FL (where N is a positive integer and FL is the frame length) is subsequently applied to the long feature vector sequence. Feature vectors within the window constitute a feature vector for an audio segment, which is used to establish a CCGMM. The window is shifting forward by r time.
To determine if a segment is substantially the same as the target audio clip, KL-max distance may be calculated between the model of the segment and the target audio clip as follows,
If the KL-max distance so calculated is below a predetermined threshold, the audio clip may be considered to be detected. As the window applied over input audio stream 520 shifts forward in time, distances typically show certain continuity from one time-step to the next. In other words, if the distance is too large, it is unlikely that one or more segments immediately following the current segment matches the target audio clip. Thus, depending on the value of the distance, a certain number of immediately following segments in the same audio stream/substream may be skipped from search.
Additionally, the partitioning mechanism partitions a large audio database into groups with proper sizes to reduce the overlapped computation (in the situation where a large audio stream is partitioned into multiple overlapped smaller substreams) and load imbalance in parallel processing by multiple processors. Smaller group size may result in large overlapped computation, while larger group size may result in considerable load imbalance. In one embodiment, the group size may be about 25 times of the size of the target audio clip.
Turning back to
I/O optimizer 630 may optimize I/O traffic on the system interconnect (e.g., system bus connecting a shared system memory with processors in the system). The I/O optimizer may decide not to load the entire audio database to be searched for from the disk into the memory in the beginning while the data range for each processor is being defined. Additionally, the I/O optimizer may let each processor read only a portion of its assigned segment from the memory at one time. By optimizing the I/O traffic, the I/O optimizer may reduce I/O contention, implement the overlap of I/O operations and computation, and help to improve computation efficiency. As a result, the scalability of audio search can be significantly improved.
Audio search module Audio 600 also comprises a plurality of audio searchers 640A through 640N. Each audio searcher (e.g., 640A) is located in a processor to process a group assigned to the processor and to search for the target audio clip. Similar to an audio searching module 400 shown in
Lines 810 through 846 illustrate how each of NG groups are processed and searched for the target in parallel by a processor in the multiprocessor system. It is worth noting that for illustration purpose, process in lines 812 to 846 is shown as iteration from the first group until the last group. In practice, if there are several processors available, several groups are processed in parallel by these available processors. At line 814, some or all of audio streams in each group may be further partitioned into NS partially overlapped segments if such streams are longer in time than the target audio clip. Line 816 starts iterative process for each segment of the group, shown in lines 818 through 832. At line 820, a feature vector sequence (frame by frame) may be extracted from the segment. At line 822, a model (e.g., CCGMM as shown in Equations (1) to (3)) may be established for the segment. At line 824, distance (e.g., KL-max distance as shown in Equation (4)) between the segment model and the target audio clip model may be computed. At line 826, whether the segment matches the target audio clip or not may be determined based on the distance calculated in line 824 and a predetermined threshold #1. If the distance is less than the threshold #1, the segment matches the target audio clip. At line 828, whether a number of following segments (e.g., M segments) in the same audio stream/substream may be skipped from searching may be determined based on the distance calculated in line 824 and a predetermined threshold #2. If the distance is larger than the threshold #2, M segments may be skipped from searching. In one embodiment, the number of segments to be skipped may vary depending upon the value of the distance. At line 830, the search results (e.g., index or starting time of a match segment in each group) may be stored in an array which is local to the processor that processes the group. At line 842, search results from local arrays from all of the processors may be summarized and outputted to a user.
Using the robust and parallel search strategy as outlined in
In one embodiment, a modified search strategy may be used. Using this strategy, a preliminary model (e.g., CCGMM) may be established for the first K frames (K>=1) of the target audio clip along with a full model for the entire target audio clip. Accordingly, a preliminary model (e.g., CCGMM) may be first established for the first K frames (K>=1) of an audio segment. During active search, the preliminary model of the first K frames of each audio segment may be first compared with the preliminary model of the first K frames of the target audio clip to produce a preliminary similarity measure. If the preliminary similarity measure indicates that these two preliminary models are significantly similar, a full model may be established for the entire audio segment and compared with the full model of the entire target audio clip; otherwise, no full model will be established for the audio segment and the next segment may be searched by first establishing a preliminary model for its first K frames and by comparing this preliminary model with the preliminary model of the target audio clip. This modified search strategy may further reduce computation load.
Although an example embodiment of the disclosed subject matter is described with reference to block and flow diagrams in
In the preceding description, various aspects of the disclosed subject matter have been described. For purposes of explanation, specific numbers, systems and configurations were set forth in order to provide a thorough understanding of the subject matter. However, it is apparent to one skilled in the art having the benefit of this disclosure that the subject matter may be practiced without the specific details. In other instances, well-known features, components, or modules were omitted, simplified, combined, or split in order not to obscure the disclosed subject matter.
Various embodiments of the disclosed subject matter may be implemented in hardware, firmware, software, or combination thereof, and may be described by reference to or in conjunction with program code, such as instructions, functions, procedures, data structures, logic, application programs, design representations or formats for simulation, emulation, and fabrication of a design, which when accessed by a machine results in the machine performing tasks, defining abstract data types or low-level hardware contexts, or producing a result.
For simulations, program code may represent hardware using a hardware description language or another functional description language which essentially provides a model of how designed hardware is expected to perform. Program code may be assembly or machine language, or data that may be compiled and/or interpreted. Furthermore, it is common in the art to speak of software, in one form or another as taking an action or causing a result. Such expressions are merely a shorthand way of stating execution of program code by a processing system which causes a processor to perform an action or produce a result.
Program code may be stored in, for example, volatile and/or non-volatile memory, such as storage devices and/or an associated machine readable or machine accessible medium including solid-state memory, hard-drives, floppy-disks, optical storage, tapes, flash memory, memory sticks, digital video disks, digital versatile discs (DVDs), etc., as well as more exotic mediums such as machine-accessible biological state preserving storage. A machine readable medium may include any mechanism for storing, transmitting, or receiving information in a form readable by a machine, and the medium may include a tangible medium through which electrical, optical, acoustical or other form of propagated signals or carrier wave encoding the program code may pass, such as antennas, optical fibers, communications interfaces, etc. Program code may be transmitted in the form of packets, serial data, parallel data, propagated signals, etc., and may be used in a compressed or encrypted format.
Program code may be implemented in programs executing on programmable machines such as mobile or stationary computers, personal digital assistants, set top boxes, cellular telephones and pagers, and other electronic devices, each including a processor, volatile and/or non-volatile memory readable by the processor, at least one input device and/or one or more output devices. Program code may be applied to the data entered using the input device to perform the described embodiments and to generate output information. The output information may be applied to one or more output devices. One of ordinary skill in the art may appreciate that embodiments of the disclosed subject matter can be practiced with various computer system configurations, including multiprocessor or multiple-core processor systems, minicomputers, mainframe computers, as well as pervasive or miniature computers or processors that may be embedded into virtually any device. Embodiments of the disclosed subject matter can also be practiced in distributed computing environments where tasks may be performed by remote processing devices that are linked through a communications network.
Although operations may be described as a sequential process, some of the operations may in fact be performed in parallel, concurrently, and/or in a distributed environment, and with program code stored locally and/or remotely for access by single or multi-processor machines. In addition, in some embodiments the order of operations may be rearranged without departing from the spirit of the disclosed subject matter. Program code may be used by or in conjunction with embedded controllers.
While the disclosed subject matter has been described with reference to illustrative embodiments, this description is not intended to be construed in a limiting sense. Various modifications of the illustrative embodiments, as well as other embodiments of the subject matter, which are apparent to persons skilled in the art to which the disclosed subject matter pertains are deemed to lie within the scope of the disclosed subject matter.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US5765166 *||Apr 23, 1996||Jun 9, 1998||Raytheon Company||Use of symmetric multiprocessors for multiple hypothesis tracking|
|US5793444 *||Jul 6, 1995||Aug 11, 1998||Lg Electronics Inc.||Audio and video signal recording and reproduction apparatus and method|
|US6181867 *||Jun 7, 1995||Jan 30, 2001||Intervu, Inc.||Video storage and retrieval system|
|US6182061 *||Apr 9, 1998||Jan 30, 2001||International Business Machines Corporation||Method for executing aggregate queries, and computer system|
|US6260036 *||May 7, 1998||Jul 10, 2001||Ibm||Scalable parallel algorithm for self-organizing maps with applications to sparse data mining problems|
|US6269376 *||Oct 26, 1998||Jul 31, 2001||International Business Machines Corporation||Method and system for clustering data in parallel in a distributed-memory multiprocessor system|
|US6381601 *||Dec 22, 1999||Apr 30, 2002||Hitachi, Ltd.||Grouping and duplicate removal method in a database|
|US6453252||May 15, 2000||Sep 17, 2002||Creative Technology Ltd.||Process for identifying audio content|
|US6460035 *||Jan 8, 1999||Oct 1, 2002||International Business Machines Corporation||Probabilistic data clustering|
|US6470331 *||Dec 4, 1999||Oct 22, 2002||Ncr Corporation||Very large table reduction in parallel processing database systems|
|US6496834 *||Dec 22, 2000||Dec 17, 2002||Ncr Corporation||Method for performing clustering in very large databases|
|US6581058 *||Mar 29, 1999||Jun 17, 2003||Microsoft Corporation||Scalable system for clustering of large databases having mixed data attributes|
|US20020129038 *||Dec 18, 2000||Sep 12, 2002||Cunningham Scott Woodroofe||Gaussian mixture models in a data mining system|
|US20030212692 *||Apr 29, 2003||Nov 13, 2003||Campos Marcos M.||In-database clustering|
|US20040002935 *||Jun 27, 2002||Jan 1, 2004||Hagai Attias||Searching multi-media databases using multi-media queries|
|US20040181523 *||Mar 23, 2004||Sep 16, 2004||Jardin Cary A.||System and method for generating and processing results data in a distributed system|
|CN1592906A||Jul 26, 2001||Mar 9, 2005||沙扎姆娱乐有限公司||System and methods for recognizing sound and music signals in high noise and distortion|
|CN1755796A||Sep 30, 2004||Apr 5, 2006||国际商业机器公司||Distance defining method and system based on statistic technology in text-to speech conversion|
|TW220483B||Title not available|
|1||European Patent Office, European Search Report mailed on Sep. 24, 2010 in E.P. patent application No. 06761349.7.|
|2||Korean Intellectual Property Office, Notice of Preliminary Rejection mailed Oct. 25, 2010 in Korean patent application number 10-2009-7000034.|
|3||Kunio Kashino, et al., "A Quick Search Method For Audio And Video Signals Based On Histogram Pruning," Sep. 2003, pp. 348-357.|
|4||O. Kao, "SMP and Cluster Architectures for Retrieval of Images in Digital Libraries," Dec. 11, 2000, pp. 1-6.|
|5||PCT/CN2006/001550 International Search Report with Written Opinion of the International Searching Authority Mailed Jan. 11, 2007.|
|6||State Intellectual Property Office, P.R. China, First Office Action issued Oct. 13, 2010 in Chinese patent application No. 200680055729.2.|
|7||*||Yih-Ru Wang and Chen-Yu Chiang. A New Common Component GMM-Based Speaker Recognition Method. In: IEEE International Conference on Acoustics, Speech, and Signal Processing, 2005, Proceedings. (ICASSP '05), pp. I-645-I-648, Mar. 23, 2005.|
|8||Yih-Ru Wang, et al., "A New Common Component GMM-Based Speaker Recognition Method," 2005 IEEE International Conference on Acoustics, Speech and Signal Processing, Mar. 18, 2005, pp. 245-248.|
|9||*||Yurong Chen, Wei Wei, Yimin Zhang. Parallel Audio Quick Search on Shared-Memory Multiprocessor Systems. http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=04228292. PDSEC-07. Proceedings of 21st International Parallel & Distributed Processing Symposium (IPDPS'07), Long Beach, CA, USA, Mar. 26-30, 2007, 6 pages.|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US20140108448 *||Mar 30, 2012||Apr 17, 2014||Intel Corporation||Multi-sensor velocity dependent context aware voice recognition and summarization|
|U.S. Classification||707/737, 707/770|
|International Classification||G06F17/30, G06F15/16, G06F7/00|
|Cooperative Classification||G06F17/30758, G06F17/30743|
|European Classification||G06F17/30U1, G06F17/30U3E|
|Mar 17, 2008||AS||Assignment|
Owner name: INTEL CORPORATION, CALIFORNIA
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:CHEN, YURONG;REEL/FRAME:020685/0194
Effective date: 20060706
|Aug 20, 2014||FPAY||Fee payment|
Year of fee payment: 4